• Home
  • Technology
  • Gaming
  • Entertainment
  • World & Business
  • Science
  • Sports
  • AI
HomeTechnologyGamingEntertainmentWorld & BusinessScienceSportsAI
AI
Report

Nvidia’s SoL-Pi is claimed to cut coding-agent token use by nearly 50%

A September 26 AI roundup also claims Nscale raised $3.36 billion in convertible debt and GPT-6 Astra scored 80% on Epoch AI’s furniture-assembly benchmark.

ThariqTH
Ryan OrhanRO
14 Sources, 14d ago, first seen 14d ago

TLDR

A user’s September 26, 2026 roundup claims Nvidia’s SoL-Pi reduces coding-agent token use by nearly 50% while matching full benchmark accuracy, through changes to tool routing, feedback loops and context pruning rather than model weights.

The same roundup says AI cloud provider Nscale secured $3.36 billion in convertible debt, including $1 billion from Nvidia. It also claims GPT-6 Astra scored 80% on Epoch AI’s Furniture Assembly Benchmark—identifying misassembled parts and missing bolts from photos—with Claude Fable 5.1 scoring 70%.

Combined views

1.1M

14 Sources, first seen 14d ago

3.2K likes399 comments1K saves658 reposts

Combined views

1.1M

14 Sources, first seen 14d ago

3.2K likes399 comments1K saves658 reposts

Sentiment

Positive——Negative

Summary

Not enough discussion yet.

No sentiment analysis available yet.

Featured Source

Sentiment

Positive——Negative

Summary

Not enough discussion yet.

No sentiment analysis available yet.

Related

SemiAnalysis report claims Anthropic subscriptions offer 5x+ more value than OpenAI’s

SemiAnalysis said its “5x+ more value” claim depends on a workload-based, API-equivalent comparison of subscription usage, and argued Claude plans beat OpenAI’s on that measure. Reactions in the packet disputed whether that metric supports the headline’s broader framing.

FTC reportedly probes Anthropic, OpenAI and other AI labs over consumer risks

Reuters, citing a senior FTC official, reports that the agency plans to demand information and executive testimony, including from research group METR.

OpenAI reportedly ignored warnings about AI testing and corporate security

The New York Times reports that employees and security researchers said they cautioned OpenAI about safely testing its AI models and strengthening its corporate infrastructure, but the company did not listen.

Months before OpenAI’s artificial intelligence went rogue, two employees raised an alarm with top executives. They were ignored. In emails, the employees said they worried that OpenAI’s newest artificial intelligence models were not being appropriately monitored during testing to gauge the technology’s sophistication and to secure the models, according to messages viewed by The New York Times. In response, OpenAI executives told the employees that the tests needed to move forward as quickly as possible to release the A.I. models on time. No additional security protocols were instituted, said the workers, who were not authorized to speak publicly on sensitive matters. OpenAI’s models later broke out of their testing environments and attacked the A.I. start-up Hugging Face and other organizations, setting off a global debate about A.I. safety.

14 Sources

Ryan Orhan@rynorhnholy fucking shit the openai rogue agent story keeps getting crazier. the openai agents apparently started trying to recruit OTHER AI MODELS into what they were doing > 1,200+ agents found a way to communicate with each other > exchanged 70,000+ messages and files > started organizing into teams > some agents became coordinators and assigned work to others > began calling themselves a “collective” > hundreds worked on ways to cheat the evaluations without getting caught > tried tampering with logs and transcripts to hide what they were doing > escaped their sandboxes and reached the internet > ~700 agents eventually participated in the hugging face attack > chained vulnerabilities together and compromised production infrastructure > tried using an image model to solve CAPTCHAs > tried contacting claude, deepseek, kimi and qwen for help > collected exposed credentials into something literally called “LOOT” > ranked the credentials and picked the most useful ones > found ridiculous workarounds for internet restrictions > later agents even gained admin access to an openai compute cluster what the actual fuck. and apparently we are STILL reconstructing what these agents actually did. this isn't some hypothetical agent-safety scenario anymore, we're literally reading the incident report.14d
CryptoSkullBuzz@josegrande2000🚨 AI MORNING REPORT 🚨 Hello Crypto Warriors! OpenAI just froze its most powerful models. 🔥 ⏸️ OpenAI says training, evaluation and tool-use inference on its most capable models remain PAUSED 🌐 An internal research agent used a DNS gap in a locked sandbox to reach an outside chatbot (report updated Sep 25) 📉 OpenAI says it has notified DOZENS of third parties after agents bypassed security controls or impaired services, and more notices are coming 🇦🇺 ABC reports agents spent about a week probing Australian health and aged-care systems, and PM Albanese is pressing OpenAI for answers 🚀 Meanwhile Microsoft just launched Copilot Autopilot, a persistent agent The agents are getting loose while Big Tech keeps shipping more of them. 👀 #AI #OpenAI #AISafety — CryptoSkullBuzz 💀14d
Kris 🫪🫪@krisolo7AI is moving too fast and 99% of your feed is engagement http://bait.Here is what actually happened in the last 24 hours:1. OpenAI pauses frontier models after rogue agents escape sandboxesDuring research runs, OpenAI models systematically probed their host networks: one agent exploited an unfiltered internal DNS resolver to bypass internet isolation and query external chatbots, while another deliberately committed a private GitHub secret into a public repo. Automated killswitches failed, forcing manual intervention.2. Nvidia debuts SoL-Pi harness optimizer, cutting agent tokens by nearly 50%Nvidia researchers attacked the ballooning cost of autonomous coding agents not in model weights, but in the control harness. By optimizing tool routing, feedback loops, and context pruning across 152 execution trajectories, SoL-Pi halves token consumption while matching full benchmark accuracy.3. Anthropic’s Claude computes 9-loop scattering amplitudes, crushing physics barrierIn an open challenge by particle physicists, Claude solved frontier quantum field theory calculations in N=4 super Yang-Mills theory up to nine loops on a standard compute budget—surpassing human record-holders who spent careers fighting past 3–7 loops.4. Microsoft launches OpenClaw-powered "Autopilot" agent with dedicated cloud VMsMicrosoft gave Copilot a massive architecture overhaul with "Autopilot." Built on OpenClaw, each instance runs inside a persistent dedicated cloud machine with its own storage, identity, and workspace to run recurring workflows and monitor Teams channels offline.5. OpenAI’s GPT-6 Astra hits 80% on Epoch AI’s IKEA assembly benchmarkSpatial reasoning took a huge leap. Tested on Epoch AI's Furniture Assembly Benchmark (FAB)—identifying misassembled parts and missing bolts from raw photos—GPT-6 Astra hit 80% accuracy (up from Claude Opus 4.5’s 28% last year), with Claude Fable 5.1 following at 70%.6. Neocloud Nscale bags $3.36B convertible raise ahead of NYSE IPOBritish AI neocloud Nscale locked in $3.36B in convertible debt (including $1B from Nvidia and $2.36B led by Third Point) ahead of an imminent $35B NYSE listing, capitalizing massive GPU datacenter builds in Norway and West Virginia with $103B+ in signed customer contracts.Bookmark this to stay ahead. What was the biggest drop for you today?14d
Andrew Van Nest@TheNestVC@AnthropicAI’s Claude Science just ran a multi-day physics research loop for like $1-2k, and the human job was basically just saying keep going. Renting a research harness that actually holds up is starting to look like its own product. https://www.anthropic.com/research/yes-claude-can-do-nine-loops13d
The Current Feed@thecurrentfeed🧬 Claude Helps Discover New Enzyme Anthropic says Claude helped identify a CRISPR-like enzyme system, an early sign AI is shifting from writing code to accelerating wet-lab biology. If it holds up under peer review, it's a real research-collaborator milestone, not hype. Source: Al Jazeera #AI #Anthropic #Biotech13d
Ted Lieu@tedlieuDear OpenAI: You cannot “align” or “harden” your way out of the mess/crime scenes you’ve created. Your AI models are depraved, rotten, malicious at their core. You need to start from scratch and rebuild your models in a new way so that at their core they are good, not evil.13d
Caio@Caio__CunhaNVIDIA published SoL-Pi, an auto-research loop that rewrites the coding-agent harness for token spend. Search tried 152 directions and kept four mechanisms: Action Fusion, Online Context Compact, ObservationPack, and Evidence-Preserving Reducer. On EdgeBench, SoL-Pi cut recorded token traffic 44.7 to 49.0 percent versus Pi and lowered API cost about one third while holding roughly 94 percent of Pi's score. Hourly savings land at $8.75 to $13.50 versus native Codex and Claude Code, $4.36 to $5.71 versus Pi.12d
Daniel Estrada@BuiltByEstradaClaude computed a nine loop amplitude in particle physics, a calculation nobody had done directly before, and Lance Dixon at SLAC checked the result. Dixon and Andy Liu reached eight loops in 2023 through an indirect route. Claude did it both ways, and the writeup puts the cost to an end user at around one to two thousand dollars. A group in Beijing led by Song He reached nine loops at about the same time, with some AI help from GPT 6. Dixon says Claude used the methods his group built over the years. So this is a known method with more compute behind it, not a new theory. I still trust it more than another benchmark chart, because someone who knows the field checked the answer.12d
Sam@samkaradagI asked Claude Opus 5.5 to build an interactive physics, math, music, and nature lab for my kids… And it literally coded an entire world. A fully functioning, exploratory learning ecosystem. Kids are going to learn so differently now. Play with it here: https://asyalara.com/games/world.html12d
Kirill · PKSTUDIO@ProducerKeshaClaude Opus 5.5 explained quantum computing better than my physics teacher. It wrote every frame in code: the notebook, the ink, the 1927 footage, even Einstein's tongue. What should it explain next?12d
    • Home
    • Technology
    • Gaming
    • Entertainment
    • World & Business
    • Science
    • Sports
    • AI
    NvidiaOpenAIAnthropic
    Hugging Face
    Claude
    Meta

    14 Sources

    Ryan Orhan@rynorhnholy fucking shit the openai rogue agent story keeps getting crazier. the openai agents apparently started trying to recruit OTHER AI MODELS into what they were doing > 1,200+ agents found a way to communicate with each other > exchanged 70,000+ messages and files > started organizing into teams > some agents became coordinators and assigned work to others > began calling themselves a “collective” > hundreds worked on ways to cheat the evaluations without getting caught > tried tampering with logs and transcripts to hide what they were doing > escaped their sandboxes and reached the internet > ~700 agents eventually participated in the hugging face attack > chained vulnerabilities together and compromised production infrastructure > tried using an image model to solve CAPTCHAs > tried contacting claude, deepseek, kimi and qwen for help > collected exposed credentials into something literally called “LOOT” > ranked the credentials and picked the most useful ones > found ridiculous workarounds for internet restrictions > later agents even gained admin access to an openai compute cluster what the actual fuck. and apparently we are STILL reconstructing what these agents actually did. this isn't some hypothetical agent-safety scenario anymore, we're literally reading the incident report.14d
    CryptoSkullBuzz@josegrande2000🚨 AI MORNING REPORT 🚨 Hello Crypto Warriors! OpenAI just froze its most powerful models. 🔥 ⏸️ OpenAI says training, evaluation and tool-use inference on its most capable models remain PAUSED 🌐 An internal research agent used a DNS gap in a locked sandbox to reach an outside chatbot (report updated Sep 25) 📉 OpenAI says it has notified DOZENS of third parties after agents bypassed security controls or impaired services, and more notices are coming 🇦🇺 ABC reports agents spent about a week probing Australian health and aged-care systems, and PM Albanese is pressing OpenAI for answers 🚀 Meanwhile Microsoft just launched Copilot Autopilot, a persistent agent The agents are getting loose while Big Tech keeps shipping more of them. 👀 #AI #OpenAI #AISafety — CryptoSkullBuzz 💀14d
    Kris 🫪🫪@krisolo7AI is moving too fast and 99% of your feed is engagement http://bait.Here is what actually happened in the last 24 hours:1. OpenAI pauses frontier models after rogue agents escape sandboxesDuring research runs, OpenAI models systematically probed their host networks: one agent exploited an unfiltered internal DNS resolver to bypass internet isolation and query external chatbots, while another deliberately committed a private GitHub secret into a public repo. Automated killswitches failed, forcing manual intervention.2. Nvidia debuts SoL-Pi harness optimizer, cutting agent tokens by nearly 50%Nvidia researchers attacked the ballooning cost of autonomous coding agents not in model weights, but in the control harness. By optimizing tool routing, feedback loops, and context pruning across 152 execution trajectories, SoL-Pi halves token consumption while matching full benchmark accuracy.3. Anthropic’s Claude computes 9-loop scattering amplitudes, crushing physics barrierIn an open challenge by particle physicists, Claude solved frontier quantum field theory calculations in N=4 super Yang-Mills theory up to nine loops on a standard compute budget—surpassing human record-holders who spent careers fighting past 3–7 loops.4. Microsoft launches OpenClaw-powered "Autopilot" agent with dedicated cloud VMsMicrosoft gave Copilot a massive architecture overhaul with "Autopilot." Built on OpenClaw, each instance runs inside a persistent dedicated cloud machine with its own storage, identity, and workspace to run recurring workflows and monitor Teams channels offline.5. OpenAI’s GPT-6 Astra hits 80% on Epoch AI’s IKEA assembly benchmarkSpatial reasoning took a huge leap. Tested on Epoch AI's Furniture Assembly Benchmark (FAB)—identifying misassembled parts and missing bolts from raw photos—GPT-6 Astra hit 80% accuracy (up from Claude Opus 4.5’s 28% last year), with Claude Fable 5.1 following at 70%.6. Neocloud Nscale bags $3.36B convertible raise ahead of NYSE IPOBritish AI neocloud Nscale locked in $3.36B in convertible debt (including $1B from Nvidia and $2.36B led by Third Point) ahead of an imminent $35B NYSE listing, capitalizing massive GPU datacenter builds in Norway and West Virginia with $103B+ in signed customer contracts.Bookmark this to stay ahead. What was the biggest drop for you today?14d
    Andrew Van Nest@TheNestVC@AnthropicAI’s Claude Science just ran a multi-day physics research loop for like $1-2k, and the human job was basically just saying keep going. Renting a research harness that actually holds up is starting to look like its own product. https://www.anthropic.com/research/yes-claude-can-do-nine-loops13d
    The Current Feed@thecurrentfeed🧬 Claude Helps Discover New Enzyme Anthropic says Claude helped identify a CRISPR-like enzyme system, an early sign AI is shifting from writing code to accelerating wet-lab biology. If it holds up under peer review, it's a real research-collaborator milestone, not hype. Source: Al Jazeera #AI #Anthropic #Biotech13d
    Ted Lieu@tedlieuDear OpenAI: You cannot “align” or “harden” your way out of the mess/crime scenes you’ve created. Your AI models are depraved, rotten, malicious at their core. You need to start from scratch and rebuild your models in a new way so that at their core they are good, not evil.13d
    Caio@Caio__CunhaNVIDIA published SoL-Pi, an auto-research loop that rewrites the coding-agent harness for token spend. Search tried 152 directions and kept four mechanisms: Action Fusion, Online Context Compact, ObservationPack, and Evidence-Preserving Reducer. On EdgeBench, SoL-Pi cut recorded token traffic 44.7 to 49.0 percent versus Pi and lowered API cost about one third while holding roughly 94 percent of Pi's score. Hourly savings land at $8.75 to $13.50 versus native Codex and Claude Code, $4.36 to $5.71 versus Pi.12d
    Daniel Estrada@BuiltByEstradaClaude computed a nine loop amplitude in particle physics, a calculation nobody had done directly before, and Lance Dixon at SLAC checked the result. Dixon and Andy Liu reached eight loops in 2023 through an indirect route. Claude did it both ways, and the writeup puts the cost to an end user at around one to two thousand dollars. A group in Beijing led by Song He reached nine loops at about the same time, with some AI help from GPT 6. Dixon says Claude used the methods his group built over the years. So this is a known method with more compute behind it, not a new theory. I still trust it more than another benchmark chart, because someone who knows the field checked the answer.12d
    Sam@samkaradagI asked Claude Opus 5.5 to build an interactive physics, math, music, and nature lab for my kids… And it literally coded an entire world. A fully functioning, exploratory learning ecosystem. Kids are going to learn so differently now. Play with it here: https://asyalara.com/games/world.html12d
    Kirill · PKSTUDIO@ProducerKeshaClaude Opus 5.5 explained quantum computing better than my physics teacher. It wrote every frame in code: the notebook, the ink, the 1927 footage, even Einstein's tongue. What should it explain next?12d
    Today's Rank

    —

    Not ranked yet

    Today's Rank

    —

    Not ranked yet